Exercise 1: Inferring Functions

Consider our model of function inference from the chapter. We can reconceptualize our program as a sequence-generator by making the input arguments 1,2,3,…. Suppose that the first number in the sequence f(1) is 1 and the second number f(2) is 4. What number would come next?

var plus = {fn: function(a,b) {return a + b}, expr: '+'}
var multiply = {fn: function(a,b) {return Math.round(a * b,0)}, expr: '*'}
var divide = {fn: function(a,b) {return Math.round(a/b,0)}, expr: '/'}
var minus = {fn: function(a,b) {return a - b}, expr: '-'}
var power = {fn: function(a,b) {return Math.pow(a,b)}, expr: '**'}
var binaryOps = [plus, multiply, divide, minus, power]

var identity = {fn: function(x) {return x}, expr: 'x'}

var randomConstantFunction = function() {
  var c = uniformDraw(_.range(10))
  return {fn: function(x){return c}, expr: c}

var randomCombination = function(f,g) {
  var op = uniformDraw(binaryOps);
  var opfn = op.fn
  var ffn = f.fn
  var gfn = g.fn
  return {fn: function(x){return opfn(ffn(x),gfn(x))}, 
          expr: f.expr+op.expr+g.expr}

// sample an arithmetic expression
var randomArithmeticExpression = function() {
  if (flip()) {
    return randomCombination(randomArithmeticExpression(), 
  } else {
    return flip() ? identity : randomConstantFunction()

viz.table(Infer({method: 'enumerate', maxExecutions: 1000}, function() {
  var e = randomArithmeticExpression();
  var f = e.fn;
  condition(f(1) == 1);
  condition(f(2) == 4);
  return f(3); // use this for Exercise 1.1
//   return e.expr; // use this for Exercise 1.2

Exercise 1.1

Not surprisingly, the model predicts 9 as the most likely result for f(3). However, it also puts significant probability on 27. Explain why these two numbers have the highest posterior probabilities.

Exercise 1.2

Why is the probability of x ** 2 is so much lower than x * x?

Exercise 1.3

Many people find the high probability assigned to 27 to be unintuitive (i.e. if we ran this as an experiment, 27 would be a very infrequent response). This suggests our model is an imperfect model of human intuitions. How could we decrease the probability of inferring 27?

HINT: Consider the priors.

Exercise 2: Role-governed concepts (optional)

In the Rational Rules model we saw in the chapter, concepts were defined in terms of the features of single objects (e.g. “it’s a raven if it has black wings”). Psychologists have suggested that many concepts are not defined by the features of a single objects, but instead by the relations the object has to other objects. For instance, “a key is something that opens a lock”. These are called role-governed concepts.

Extend the Rational Rules model to capture role-governed concepts.

Hint: You will need primitive relations in your language of thought.

Hint: Consider adding quantifiers (e.g. there exists) to your language of thought.